Methodology
Offsend Radar publishes weekly, public-safe aggregates about AI-context hygiene in open-source repositories. This page explains what we scan, what we exclude, and how results reach the dashboard.
What we scan
Each week the orchestrator clones target repositories and runsoffsend report — the same public-safe mode available in the Offsend CLI. Checks are path-level only:
- Exposed patterns — sensitive file categories (e.g.
.env, key material) visible at the path level - Ignore file coverage — presence of AI-oriented ignore files (
.cursorignore,.claudeignore, project rules, and similar)
Scans use shallow clones of a single ref per repository from the weekly cohort list.
What we do not scan or publish
- File contents, secret values, tokens, or credentials
- Exact file paths or filenames in public exports
- Repository names alongside findings unless the maintainer opted in
- Claims of security, compliance, or “secret-free” status
Private orchestration metadata stays in the privateradar-state repository and is never copied to public exports — except for opted-in named listings.
How repositories are selected
Weekly cohorts are defined manually in radar-state/targets/<week>.yml — typically open-source repositories the Offsend team tracks or that maintainers opt into. Repos can be skipped when archived or deprecated.
Named public listings
Repository names appear on this site only when a maintainer has consented viaradar-state/registry/public-listings.yml. Listings use positive framing (e.g. “AI Context Reviewed”), not blame-oriented labels. Fleet aggregates remain anonymous.
How results are sanitized
offsend reportproduces anonymized per-repo JSON (pattern IDs and counts, no paths).- Fleet aggregates roll up counts across repos in the private data store.
export/<week>/summary.jsonis the public-safe subset — fleet aggregates plus opted-in listings.- This site renders only the export data.
Reproduce checks locally
brew install --cask offsend/tap/offsend-cli
offsend show
offsend report . --out report.jsonEverything runs locally. No file contents are uploaded.
Principles
- Path-level by default — structure and ignore files, not file contents
- No secrets collection — reports must not contain sensitive values
- No public shaming — aggregate patterns, not individual blame
- Opt-in for names — maintainers choose whether to appear by name
- No security claims — hygiene signals only, not certification